Next-Gen Dynamic Hand Gesture Recognition: MediaPipe, Inception-v3 and LSTM-Based Enhanced Deep Learning Model

Abstract

Gesture recognition is crucial in computer vision-based applications, such as drone control, gaming, virtual and augmented reality (VR/AR), and security, especially in human–computer interaction (HCI)-based systems. There are two types of gesture recognition systems, i.e., static and dynamic. However, our focus in this paper is on dynamic gesture recognition. In dynamic hand gesture recognition systems, the sequences of frames, i.e., temporal data, pose significant processing challenges and reduce efficiency compared to static gestures. These data become multi-dimensional compared to static images because spatial and temporal data are being processed, which demands complex deep learning (DL) models with increased computational costs. This article presents a novel triple-layer algorithm that efficiently reduces the 3D feature map into 1D row vectors and enhances the overall performance. First, we process the individual images in a given sequence using the MediaPipe framework and extract the regions of interest (ROI). The processed cropped image is then passed to the Inception-v3 for the 2D feature extractor. Finally, a long short-term memory (LSTM) network is used as a temporal feature extractor and classifier. Our proposed method achieves an average accuracy of more than 89.7%. The experimental results also show that the proposed framework outperforms existing state-of-the-art methods.